Dmytro Kovalskyi
Dmytro Kovalskyi
Please keep in mind that even if local input helps, we have no space to copy the streamer file on a worker node - we have limited amount of space...
On the other hand, our biggest problem is that all these slow jobs create manual work for operators. If we have a procedure to retry that has a better success...
Danilo summarized it well. We were always using cores as a proxy for the scratch space assuming we have 20GB/core.
We compute how much memory we need and then based on that request how many cores we require. This was introducing significant inefficiency in repacking since the second core was...
I think we mostly need to fix L1SCOUT. It's completely wrong as far as I understand and it needs to be updated/expanded.
We don't run reco of any sort on HLTSCOUT, just repacking, which keeps everything included by TSG in the stream.
I would suggest to keep code expecting non-standard input data format separately from the prompt reco that takes RAW data as input, i.e. `hltScouting_Run3_2024::promptReco` seems like a better option to...
I leave it to @germanfgv to check it. We will need to do a proper test with a Tier-0 replay to make sure everything works properly when we have this...
Someone needs to look at this at high level. I can help if the issue is with XGBoost and it's C API itself. It's not fun, but I had to...
@ftorrresd, we will have a new era after special heavy ion runs. Previous releases are not going to be used for production or data reprocessing involving BPHNano.